openstef 4.1.2__tar.gz → 4.2.2__tar.gz

{openstef-4.1.2 → openstef-4.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: openstef
-Version: 4.1.2
+Version: 4.2.2
 Summary: Open Short Term Energy forecasting library
 Project-URL: Documentation, https://openstef.github.io/openstef/index.html
 Project-URL: Homepage, https://lfenergy.org/projects/openstef/
@@ -16,26 +16,26 @@ Classifier: Programming Language :: Python :: 3 :: Only
 Classifier: Programming Language :: Python :: 3.12
 Classifier: Programming Language :: Python :: 3.13
 Requires-Python: <4.0,>=3.12
-Requires-Dist: openstef-core==4.1.2
-Requires-Dist: openstef-models[xgb-cpu]==4.1.2
+Requires-Dist: openstef-core==4.2.2
+Requires-Dist: openstef-models[xgb-cpu]==4.2.2
 Provides-Extra: all
-Requires-Dist: openstef-beam[all]==4.1.2; extra == 'all'
-Requires-Dist: openstef-core==4.1.2; extra == 'all'
-Requires-Dist: openstef-foundation-models[cpu]==4.1.2; extra == 'all'
-Requires-Dist: openstef-meta==4.1.2; extra == 'all'
-Requires-Dist: openstef-models[xgb-cpu]==4.1.2; extra == 'all'
+Requires-Dist: openstef-beam[all]==4.2.2; extra == 'all'
+Requires-Dist: openstef-core==4.2.2; extra == 'all'
+Requires-Dist: openstef-foundation-models[cpu]==4.2.2; extra == 'all'
+Requires-Dist: openstef-meta==4.2.2; extra == 'all'
+Requires-Dist: openstef-models[xgb-cpu]==4.2.2; extra == 'all'
 Provides-Extra: beam
 Requires-Dist: huggingface-hub>=0.35.3; extra == 'beam'
-Requires-Dist: openstef-beam==4.1.2; extra == 'beam'
+Requires-Dist: openstef-beam==4.2.2; extra == 'beam'
 Provides-Extra: core
-Requires-Dist: openstef-core==4.1.2; extra == 'core'
+Requires-Dist: openstef-core==4.2.2; extra == 'core'
 Provides-Extra: foundation-models
-Requires-Dist: openstef-foundation-models[cpu]==4.1.2; extra == 'foundation-models'
+Requires-Dist: openstef-foundation-models[cpu]==4.2.2; extra == 'foundation-models'
 Provides-Extra: meta
-Requires-Dist: openstef-meta==4.1.2; extra == 'meta'
+Requires-Dist: openstef-meta==4.2.2; extra == 'meta'
 Provides-Extra: models
-Requires-Dist: openstef-meta==4.1.2; extra == 'models'
-Requires-Dist: openstef-models[xgb-cpu]==4.1.2; extra == 'models'
+Requires-Dist: openstef-meta==4.2.2; extra == 'models'
+Requires-Dist: openstef-models[xgb-cpu]==4.2.2; extra == 'models'
 Description-Content-Type: text/markdown
 
 <!--

openstef-4.2.2/docs/automation.yml

@@ -0,0 +1,56 @@
+# SPDX-FileCopyrightText: 2026 Contributors to the OpenSTEF project <openstef@lfenergy.org>
+#
+# SPDX-License-Identifier: MPL-2.0
+
+# docs/automation.yml — last-match-wins mode rules
+#
+# Determines how the pipeline treats each docs file.
+# Rules are evaluated in order; the LAST matching rule wins.
+#
+# Modes:
+#   manual     — never touch (human-owned pages: index, api, contribute)
+#   edit       — patch on relevant changes (existing content-rich pages)
+#   regenerate — always rebuild from scratch (requires a brief in pages.yml)
+
+rules:
+  # Default: all RST and Markdown files are editable
+  - pattern: "**/*.rst"
+    mode: edit
+  - pattern: "**/*.md"
+    mode: edit
+
+  # README is explicitly governed — change mode here to edit/regenerate/manual
+  - pattern: "README.md"
+    mode: edit
+
+  # Notebooks are human-authored — never touch
+  - pattern: "**/*.ipynb"
+    mode: manual
+
+  # API reference is auto-generated separately — never touch
+  - pattern: "docs/source/api/**"
+    mode: manual
+
+  # Contribute section is human-owned — never touch
+  - pattern: "docs/source/contribute/**"
+    mode: manual
+
+  # Top-level index is human-owned — never touch
+  - pattern: "docs/source/index.rst"
+    mode: manual
+
+  # Sphinx templates are internal build files — never touch
+  - pattern: "docs/source/_templates/**"
+    mode: manual
+
+  # Section index stubs are navigation scaffolding — human-owned
+  - pattern: "docs/source/**/index.rst"
+    mode: manual
+
+  # Project meta pages (license, maintainers, committee, etc.) are static — never touch
+  - pattern: "docs/source/project/**"
+    mode: manual
+
+  # # Installation guide is always regenerated from pyproject.toml
+  # - pattern: "docs/source/user_guide/getting_started/installation.rst"
+  #   mode: regenerate

openstef-4.2.2/docs/pages.yml

@@ -0,0 +1,310 @@
+# SPDX-FileCopyrightText: 2026 Contributors to the OpenSTEF project <openstef@lfenergy.org>
+#
+# SPDX-License-Identifier: MPL-2.0
+
+# docs/pages.yml — catalog of pages the pipeline writes
+#
+# Every page the pipeline is allowed to create or edit must be listed here.
+# Pages not listed are treated as untracked (soft flag, not blocked).
+# Pages listed as manual in automation.yml are excluded automatically.
+#
+# Fields:
+#   path        — repo-relative path to the RST file
+#   title       — canonical page title (used to normalise the RST heading)
+#   page_type   — sub_page | section_index | standalone (default: sub_page)
+#   purpose     — what this page should explain (passed to LLM as generation brief)
+#   diagrams    — list of Mermaid diagram descriptions to auto-generate (optional)
+
+pages:
+  # ── Repository README ──────────────────────────────────────────────────────
+
+  - path: README.md
+    title: OpenSTEF
+    page_type: standalone
+    purpose: >
+      The front door of the project. Keep it short and scannable (under two
+      minutes to read). Required sections in order: (1) Logo at the top.
+      (2) Status badges (build, license, version). (3) What is OpenSTEF —
+      max 4 sentences, link to the OpenSTEF website. (4) Brief monorepo
+      overview — how the repository is structured. (5) How to Install —
+      copy-paste-ready pip commands. (6) Examples — link to the examples/
+      folder. (7) License — state license, link to LICENSE file. (8)
+      Contributing — link to .github contributing pages, do not duplicate
+      inline. (9) Citations — BibTeX entry or DOI link. (10) Contact —
+      link to support page. Design principle: link out, do not duplicate
+      content that lives in the docs.
+
+  # ── Getting Started ────────────────────────────────────────────────────────
+
+  - path: docs/source/user_guide/getting_started/migration.rst
+    title: Migrating from OpenSTEF 3
+    page_type: sub_page
+    purpose: >
+      Help v3 users understand what changed and how to adapt their code.
+      v3 was a single `openstef` package with pipeline/task API plus `openstef-dbc`
+      for database access. v4 is a multi-package workspace that separates concerns.
+      Cover: (1) Why the split happened: core types vs models vs evaluation vs
+      orchestration are now independent. (2) Configuration: `PredictionJobDataClass`
+      became `ForecastingWorkflowConfig` (a full Pydantic model). Show a before/after
+      snippet. (3) Data: raw DataFrames became `TimeSeriesDataset`. (4) API: the
+      pipeline/task pattern became Workflows where users own I/O. (5) Database
+      connector: `openstef-dbc` has no v4 equivalent — users write their own data
+      integration. (6) Model type renames: 'xgb' → 'xgboost', 'lgb' → 'lgbm'.
+      (7) A migration checklist. Package mapping: openstef (core ML) → openstef-models,
+      openstef (data classes) → openstef-core, openstef (evaluation) → openstef-beam,
+      (new) → openstef-meta, openstef-dbc → no equivalent.
+    diagrams:
+      - >
+        Left-to-right graph showing v3 monolith splitting into v4 packages.
+        The v3 'openstef' box has arrows to openstef-models (Core ML), openstef-core
+        (Data Classes), openstef-beam (Evaluation), openstef-meta (New).
+        A separate 'openstef-dbc' box has an arrow to 'Your own I/O layer'
+        with a note 'v4 does not include a database connector'.
+
+  - path: docs/source/user_guide/getting_started/installation.rst
+    title: Installation
+    page_type: sub_page
+    purpose: >
+      Step-by-step installation guide generated from pyproject.toml.
+      pip commands for each package (openstef-core, openstef-models, openstef-beam,
+      openstef-meta), optional extras, and a quick-start snippet showing the
+      minimal working import. Note the meta-package openstef installs all sub-packages.
+
+  - path: docs/source/user_guide/getting_started/quick_start_tutorial.ipynb
+    title: Quick Start
+    page_type: sub_page
+    purpose: >
+      Train your first model and make a forecast in under 5 minutes.
+      This notebook is human-authored and managed manually — listed here
+      so neighbouring pages know it exists and can reference it without
+      duplicating its content.
+
+  - path: docs/source/user_guide/getting_started/foundation_model_forecasting_quickstart.ipynb
+    title: Foundation Model Quickstart
+    page_type: sub_page
+    purpose: >
+      Forecast with a pretrained foundation model in a few minutes — no
+      per-target training required. This notebook is human-authored and
+      managed manually — listed here so neighbouring pages know it exists
+      and can reference it without duplicating its content.
+
+  # ── Concepts ───────────────────────────────────────────────────────────────
+
+  - path: docs/source/user_guide/concepts/intro_to_energy_forecasting.rst
+    title: Intro to Energy Forecasting
+    page_type: sub_page
+    purpose: >
+      Context for the rest of the docs. What is short-term load forecasting and
+      why do grid operators need it? What input signals matter (load history, weather
+      forecasts, calendar features, market prices) and why each contributes. What
+      'horizon' means (15 minutes to 48 hours) and why accuracy degrades with lead time.
+      Practical difficulties: unpredictable consumers, aggregation level effects,
+      capacity changes. Keep short and focused. No diagram needed.
+
+  - path: docs/source/user_guide/concepts/models.rst
+    title: Models
+    page_type: sub_page
+    purpose: >
+      Overview of available models with selection guidance. Explain the containment
+      hierarchy: Forecaster (pure ML predictor) → ForecastingModel (adds pre/post
+      TransformPipelines) → CustomForecastingWorkflow (adds callbacks, model_id,
+      lifecycle) → Presets (opinionated factories via ForecastingWorkflowConfig).
+      Then a model-selection comparison table: Ensemble (preferred, tree+linear),
+      GBLinear (congestion management, extrapolation), LGBMLinear (balance),
+      XGBoost/LGBM (general-purpose), MedianForecaster (stable loads baseline),
+      FlatlinerForecaster and ConstantQuantileForecaster (fallback). All support
+      quantile forecasting. Link to API reference per model.
+    diagrams:
+      - >
+        Nested boxes showing containment (innermost to outermost): Forecaster
+        (pure ML predictor, e.g., XGBoostForecaster) is innermost.
+        ForecastingModel wraps Forecaster and adds preprocessing and postprocessing
+        TransformPipelines. CustomForecastingWorkflow wraps ForecastingModel and
+        adds callbacks, model_id, and lifecycle management. Outside, Presets
+        (ForecastingWorkflowConfig + create_forecasting_workflow factory) construct
+        a fully-wired Workflow from configuration. No data flow arrows — this is
+        purely about what-contains-what.
+
+  - path: docs/source/user_guide/concepts/metalearning.rst
+    title: Metalearning
+    page_type: sub_page
+    purpose: >
+      Explain metalearning as implemented in openstef-meta. Key insight: ensemble
+      models combine complementary strengths — tree-based models capture non-linear
+      patterns but cannot extrapolate; linear models can extrapolate but not learn
+      non-linear patterns. Architecture: EnsembleForecastingModel fans out to N
+      parallel per-forecaster pipelines. ForecastCombiner produces the final forecast.
+      Two strategies: WeightsCombiner (LGBM classifier per quantile, soft probability
+      weighting) and StackingCombiner (meta-regressor takes base predictions as input).
+      Link to ensemble_forecasting tutorial.
+    diagrams:
+      - >
+        EnsembleForecastingModel prediction flow: TimeSeriesDataset → shared
+        preprocessing TransformPipeline → fans out to N parallel paths
+        [per-forecaster preprocessing → Forecaster] for each base model →
+        merge into EnsembleForecastDataset → ForecastCombiner →
+        final ForecastDataset.
+      - >
+        Two formula-style diagrams side by side. WeightsCombiner: base predictions →
+        LGBM classifier per quantile → predict_proba → soft probability weights →
+        weighted average = final forecast. StackingCombiner: base predictions per
+        quantile (as input features) → meta-regressor → direct final quantile
+        prediction.
+
+  - path: docs/source/user_guide/concepts/beam.rst
+    title: BEAM
+    page_type: sub_page
+    purpose: >
+      Explain BEAM (Backtesting, Evaluation, Analysis, and Metrics). BenchmarkPipeline
+      orchestrates sequential stages per target: (A) BacktestPipeline steps through
+      data via BacktestEventGenerator, using BacktestForecasterMixin with
+      RestrictedHorizonVersionedTimeSeries to prevent data leakage. Output:
+      predictions as TimeSeriesDataset. (B) EvaluationPipeline segments by
+      available_at x lead_time x time_windows, computes configurable metrics.
+      Output: EvaluationReport. (C) AnalysisPipeline produces HTML visualisations.
+      BenchmarkStorage persists outputs. Value: fair, reproducible model comparison.
+      Link to backtesting guide for practical usage.
+    diagrams:
+      - >
+        BenchmarkPipeline as outer orchestrator containing the sequential per-target
+        flow: TargetProvider feeds data → ForecasterFactory creates BacktestForecasterMixin
+        → BacktestPipeline produces predictions (TimeSeriesDataset) →
+        EvaluationPipeline produces EvaluationReport → AnalysisPipeline produces
+        visualisations (AnalysisOutput). BenchmarkStorage persists after each stage.
+
+  - path: docs/source/user_guide/concepts/component_splitting.rst
+    title: Component Splitting
+    page_type: sub_page
+    purpose: >
+      Explain what component splitting is and why operators need it. A grid operator
+      sees aggregated load but needs to understand the constituents (solar, wind,
+      base industrial, residential). Component splitting decomposes an aggregated
+      forecast into constituents using known profiles or ratios. Critical: this does
+      NOT improve forecast accuracy — the forecast is made on aggregated load first,
+      then decomposed. Use cases: grid planning, congestion analysis, understanding
+      generation mix, regulatory reporting.
+
+  - path: docs/source/user_guide/concepts/foundation_models.rst
+    title: Foundation Models
+    page_type: sub_page
+    purpose: >
+      Zero-shot pretrained forecasting on an ONNX runtime — no per-target
+      training required. Explain what a foundation model is in the OpenSTEF
+      context, how the ONNX runtime is used for inference, what fields control model 
+      selection and execution providers (CPU, CUDA, CoreML), and how batching works 
+      across multiple targets. Cover the tradeoff vs trained models: faster 
+      deployment, no training data needed, but less adaptation to local patterns. 
+
+  # ── Guides ─────────────────────────────────────────────────────────────────
+
+  - path: docs/source/user_guide/guides/forecasting.rst
+    title: Forecasting
+    page_type: sub_page
+    purpose: >
+      The central how-to-forecast page — architecture and design decisions, not
+      runnable code (tutorials cover that). What OpenSTEF needs: datetime-indexed
+      data with a 'load' column, weather features, declared sample_interval.
+      What it returns: ForecastDataset with point forecast and optional quantiles.
+      Recommend Presets for production, Workflow for research. Data requirements:
+      expected columns, time zones, sample_interval. Fit/predict lifecycle at
+      Workflow level: fit (preprocessing → forecaster.fit → callbacks),
+      predict (preprocessing → forecaster.predict → postprocessing → callbacks).
+      Model reuse: if a recent model exists in storage, workflow skips retraining.
+      Link to forecasting_quickstart and custom_pipeline tutorials.
+    diagrams:
+      - >
+        Two parallel flows. Fit path: Historical TimeSeriesDataset →
+        preprocessing TransformPipeline (validation → feature engineering →
+        standardisation) → ForecastInputDataset → Forecaster.fit() → trained model
+        → callbacks (MLflow storage). Predict path: New TimeSeriesDataset → same
+        preprocessing → ForecastInputDataset → Forecaster.predict() → raw
+        ForecastDataset → postprocessing TransformPipeline (quantile sorting,
+        confidence intervals) → final ForecastDataset → callbacks (logging).
+        Workflow wraps both paths.
+
+  - path: docs/source/user_guide/guides/feature_engineering_tutorial.ipynb
+    title: Feature Engineering
+    page_type: sub_page
+    purpose: >
+      Create, select, and manage features for your models.
+
+  - path: docs/source/user_guide/guides/datasets_tutorial.ipynb
+    title: Datasets
+    page_type: sub_page
+    purpose: >
+      Prepare and validate input data for training and prediction.
+
+  - path: docs/source/user_guide/guides/probabilistic_forecasting.rst
+    title: Probabilistic Forecasting
+    page_type: sub_page
+    purpose: >
+      Why: a point forecast is not enough — grid operators need uncertainty ranges
+      for reserve planning, congestion management, and market bidding. How OpenSTEF
+      provides this: all models support quantile forecasting via the quantiles field
+      on ForecastingWorkflowConfig. XGBoost uses pinball loss, GBLinear uses
+      reg:quantileerror, LGBM uses MultiQuantileRegressor. Isotonic calibration:
+      IsotonicQuantileCalibrator is a postprocessing Transform that maps predicted
+      values to observed quantile levels using sklearn IsotonicRegression, restoring
+      proper coverage while enforcing monotonicity. Link to quantile_calibration
+      tutorial and IsotonicQuantileCalibrator API reference.
+
+  - path: docs/source/user_guide/guides/reliability_fallback.rst
+    title: Reliability and Fallback
+    page_type: sub_page
+    purpose: >
+      Operational reality: data feeds fail — meters flatline, weather APIs go down,
+      data arrives late. The system must detect and respond gracefully. Validation
+      transforms: FlatlineChecker (configurable via flatliner_threshold,
+      detect_non_zero_flatliner), CompletenessChecker (completeness_threshold).
+      Fallback strategies: FlatlinerForecaster (constant prediction on flatline),
+      ConstantQuantileForecaster (minimal forecast on low data availability,
+      configured via completeness_threshold_target_constant_quantile). All configured
+      in ForecastingWorkflowConfig; the preset factory switches automatically based
+      on data quality checks.
+
+  - path: docs/source/user_guide/guides/backtesting.ipynb
+    title: Backtesting
+    page_type: sub_page
+    purpose: >
+      Practical guide: how to evaluate your model on historical data before trusting
+      it in production. Configure BacktestForecasterMixin (RestrictedHorizonVersionedTimeSeries
+      prevents lookahead). Run BacktestPipeline stepping through history via
+      BacktestEventGenerator (BacktestConfig with predict_interval and train_interval).
+      Then EvaluationPipeline segments by available_at x lead_time x time_windows.
+      Interpreting results: per-window vs aggregated metrics, what good looks like
+      for energy forecasting. Link to BEAM concept page and backtesting_quickstart
+      tutorial.
+
+  - path: docs/source/user_guide/guides/deployment.rst
+    title: Deployment
+    page_type: sub_page
+    purpose: >
+      How the library works in different deployment contexts without prescribing a
+      single approach. OpenSTEF is a pure ML library with no built-in scheduling,
+      orchestration, or data storage. Three common strategies: (1) Scheduled notebooks
+      (training + forecasting notebooks with data integration, run on Sagemaker /
+      Databricks). (2) DAG-based orchestration (Airflow, Dagster, Prefect — separate
+      data integration, training, prediction stages). (3) Queued execution (per-location
+      tasks dispatched to workers, optimised for large-scale operations). Data sources:
+      link Open-Meteo for weather (free, global), KNMI for NL-specific weather. Model
+      storage: MLflow model registry via MLFlowStorageCallback handles versioning and
+      reuse. Storage is pluggable — users implement custom callbacks. OpenSTEF is
+      modular: custom callbacks, storage backends, and integration layers are
+      encouraged.
+
+  - path: docs/source/user_guide/guides/foundation_model_forecasting.ipynb
+    title: Foundation Model Forecasting
+    page_type: sub_page
+    purpose: >
+      Checkpoints, execution providers for your hardware, batching, and
+      backtesting with a pretrained foundation model. 
+
+  - path: docs/source/user_guide/logging.rst
+    title: Logging
+    page_type: sub_page
+    purpose: >
+      How to configure logging for OpenSTEF applications. Cover the default
+      NullHandler behaviour, the logger naming hierarchy and what each level
+      emits, common configuration patterns for development and production
+      (including structured JSON output), and troubleshooting tips for missing
+      or excessive output.

{openstef-4.1.2 → openstef-4.2.2}/packages/openstef-beam/pyproject.toml

@@ -7,7 +7,7 @@ requires = [ "hatchling" ]
 
 [project]
 name = "openstef-beam"
-version = "4.1.2"
+version = "4.2.2"
 description = "Backtesting, Evaluation, Analysis and Metrics (BEAM) library for OpenSTEF"
 readme = "README.md"
 keywords = [ "energy", "forecasting", "machinelearning" ]

{openstef-4.1.2 → openstef-4.2.2}/packages/openstef-core/pyproject.toml

@@ -7,7 +7,7 @@ requires = [ "hatchling" ]
 
 [project]
 name = "openstef-core"
-version = "4.1.2"
+version = "4.2.2"
 description = "Core functionality for OpenSTEF, a framework for short-term energy forecasting."
 readme = "README.md"
 keywords = [ "energy", "forecasting", "machinelearning" ]

{openstef-4.1.2 → openstef-4.2.2}/packages/openstef-foundation-models/pyproject.toml

@@ -7,7 +7,7 @@ requires = [ "hatchling" ]
 
 [project]
 name = "openstef-foundation-models"
-version = "4.1.2"
+version = "4.2.2"
 description = "Foundation model support for OpenSTEF"
 readme = "README.md"
 keywords = [ "energy", "forecasting", "foundationmodels", "machinelearning", "onnx" ]

{openstef-4.1.2 → openstef-4.2.2}/packages/openstef-meta/pyproject.toml

@@ -7,7 +7,7 @@ requires = [ "hatchling" ]
 
 [project]
 name = "openstef-meta"
-version = "4.1.2"
+version = "4.2.2"
 description = "Meta models for OpenSTEF"
 readme = "README.md"
 keywords = [ "energy", "forecasting", "machinelearning" ]

{openstef-4.1.2 → openstef-4.2.2}/packages/openstef-models/pyproject.toml

@@ -7,7 +7,7 @@ requires = [ "hatchling" ]
 
 [project]
 name = "openstef-models"
-version = "4.1.2"
+version = "4.2.2"
 description = "Core models for OpenSTEF"
 readme = "README.md"
 keywords = [ "energy", "forecasting", "machinelearning" ]

{openstef-4.1.2 → openstef-4.2.2}/packages/openstef-models/src/openstef_models/models/forecasting/lgbm_forecaster.py

@@ -111,6 +111,12 @@ class LGBMHyperParams(HyperParams):
         description="Fraction of features used when constructing each tree. Range: (0,1]",
     )
 
+    # Early Stopping
+    early_stopping_rounds: int | None = Field(
+        default=None,
+        description="Training will stop if performance doesn't improve for this many rounds. Requires validation data.",
+    )
+
     @classmethod
     def forecaster_class(cls) -> "type[LGBMForecaster]":
         """Create a LightGBM forecaster instance from this configuration.
@@ -184,10 +190,6 @@ class LGBMForecaster(Forecaster, ExplainableForecaster, ContributionsMixin):
         alias="seed",
         description="Random seed for reproducibility.",
     )
-    early_stopping_rounds: int | None = Field(
-        default=None,
-        description="Training stops if performance doesn't improve for this many rounds.",
-    )
 
     _lgbm_model: MultiQuantileRegressor = PrivateAttr()
 
@@ -229,7 +231,7 @@ class LGBMForecaster(Forecaster, ExplainableForecaster, ContributionsMixin):
             "colsample_bytree": self.hyperparams.colsample_bytree,
             # General parameters
             "random_state": self.random_state,
-            "early_stopping_rounds": self.early_stopping_rounds,
+            "early_stopping_rounds": self.hyperparams.early_stopping_rounds,
             "verbosity": self.verbosity,
             "n_jobs": self.n_jobs,
         }

{openstef-4.1.2 → openstef-4.2.2}/packages/openstef-models/src/openstef_models/models/forecasting/lgbmlinear_forecaster.py

@@ -111,6 +111,12 @@ class LGBMLinearHyperParams(HyperParams):
         description="Fraction of features used when constructing each tree. Range: (0,1]",
     )
 
+    # Early Stopping
+    early_stopping_rounds: int | None = Field(
+        default=None,
+        description="Training will stop if performance doesn't improve for this many rounds. Requires validation data.",
+    )
+
     @classmethod
     def forecaster_class(cls) -> "type[LGBMLinearForecaster]":
         """Get forecaster class for these hyperparams.
@@ -185,10 +191,6 @@ class LGBMLinearForecaster(Forecaster, ExplainableForecaster, ContributionsMixin
         alias="seed",
         description="Random seed for reproducibility.",
     )
-    early_stopping_rounds: int | None = Field(
-        default=None,
-        description="Training stops if performance doesn't improve for this many rounds.",
-    )
 
     _lgbmlinear_model: MultiQuantileRegressor = PrivateAttr()
 
@@ -230,7 +232,7 @@ class LGBMLinearForecaster(Forecaster, ExplainableForecaster, ContributionsMixin
             "colsample_bytree": self.hyperparams.colsample_bytree,
             # General parameters
             "random_state": self.random_state,
-            "early_stopping_rounds": self.early_stopping_rounds,
+            "early_stopping_rounds": self.hyperparams.early_stopping_rounds,
             "verbosity": self.verbosity,
             "n_jobs": self.n_jobs,
         }

{openstef-4.1.2 → openstef-4.2.2}/packages/openstef-models/src/openstef_models/utils/multi_quantile_regressor.py

@@ -84,6 +84,9 @@ class MultiQuantileRegressor(BaseEstimator, RegressorMixin):
         """
         # Pass model-specific eval arguments
         kwargs = {}
+
+        x_array = np.asarray(X)
+
         for model in self._models:
             # Check if early stopping is supported
             # Check that eval_set is supported
@@ -91,21 +94,20 @@ class MultiQuantileRegressor(BaseEstimator, RegressorMixin):
                 model.set_params(early_stopping_rounds=None)
 
             if eval_set is not None and self.learner_eval_sample_weight_param is not None:
+                kwargs["eval_set"] = [
+                    (x_array if eval_x is X else np.asarray(eval_x), eval_y) for eval_x, eval_y in eval_set
+                ]
                 kwargs[self.learner_eval_sample_weight_param] = eval_sample_weight
-
-            if "early_stopping_rounds" in self.hyperparams and self.learner_eval_sample_weight_param is not None:
-                model.set_params(early_stopping_rounds=self.hyperparams["early_stopping_rounds"])
+                if "early_stopping_rounds" in self.hyperparams:
+                    model.set_params(early_stopping_rounds=self.hyperparams["early_stopping_rounds"])
 
             if feature_name:
                 self.model_feature_names = feature_name
             else:
                 self.model_feature_names = []
 
-            if eval_sample_weight is not None and self.learner_eval_sample_weight_param:
-                kwargs[self.learner_eval_sample_weight_param] = eval_sample_weight
-
             model.fit(  # type: ignore
-                X=np.asarray(X),
+                X=x_array,
                 y=y,
                 sample_weight=sample_weight,
                 **kwargs,

{openstef-4.1.2 → openstef-4.2.2}/packages/openstef-models/tests/unit/utils/test_multi_quantile_regressor.py

@@ -117,3 +117,28 @@ def test_is_fitted_true_after_fit(dataset: tuple[pd.DataFrame, pd.Series], basel
 
     # Assert
     assert model.is_fitted
+
+
+def test_fit_forwards_eval_set_for_early_stopping_lgbm(dataset: tuple[pd.DataFrame, pd.Series]):
+    # Arrange
+    quantiles = [Quantile(0.1), Quantile(0.5), Quantile(0.9)]
+    early_stopping_rounds = 3
+    X, y = dataset
+    y = y.to_numpy()
+    X_train, y_train = X.iloc[:80], y[:80]  # noqa: N806
+    X_val, y_val = X.iloc[80:], y[80:]  # noqa: N806
+    model = MultiQuantileRegressor(
+        base_learner=LGBMRegressor,
+        quantile_param="alpha",
+        quantiles=quantiles,
+        hyperparams={"objective": "quantile", "n_estimators": 10, "early_stopping_rounds": early_stopping_rounds},
+    )
+
+    # Act
+    model.fit(X_train, y_train, eval_set=[(X_train, y_train), (X_val, y_val)])
+
+    # Assert
+    for estimator in model.models:
+        assert isinstance(estimator, LGBMRegressor)
+        assert estimator.get_params()["early_stopping_rounds"] == early_stopping_rounds
+        assert estimator.evals_result_

{openstef-4.1.2 → openstef-4.2.2}/pyproject.toml

@@ -7,7 +7,7 @@ requires = [ "hatchling" ]
 
 [project]
 name = "openstef"
-version = "4.1.2"
+version = "4.2.2"
 description = "Open Short Term Energy forecasting library"
 readme = "README.md"
 keywords = [ "energy", "forecasting", "machinelearning" ]
@@ -24,32 +24,32 @@ classifiers = [
   "Programming Language :: Python :: 3.13",
 ]
 dependencies = [
-  "openstef-core==4.1.2",
-  "openstef-models[xgb-cpu]==4.1.2",
+  "openstef-core==4.2.2",
+  "openstef-models[xgb-cpu]==4.2.2",
 ]
 optional-dependencies.all = [
-  "openstef-beam[all]==4.1.2",
-  "openstef-core==4.1.2",
-  "openstef-foundation-models[cpu]==4.1.2",
-  "openstef-meta==4.1.2",
-  "openstef-models[xgb-cpu]==4.1.2",
+  "openstef-beam[all]==4.2.2",
+  "openstef-core==4.2.2",
+  "openstef-foundation-models[cpu]==4.2.2",
+  "openstef-meta==4.2.2",
+  "openstef-models[xgb-cpu]==4.2.2",
 ]
 optional-dependencies.beam = [
   "huggingface-hub>=0.35.3",
-  "openstef-beam==4.1.2",
+  "openstef-beam==4.2.2",
 ]
 optional-dependencies.foundation-models = [
-  "openstef-foundation-models[cpu]==4.1.2",
+  "openstef-foundation-models[cpu]==4.2.2",
 ]
 optional-dependencies.meta = [
-  "openstef-meta==4.1.2",
+  "openstef-meta==4.2.2",
 ]
 optional-dependencies.models = [
-  "openstef-meta==4.1.2",
-  "openstef-models[xgb-cpu]==4.1.2",
+  "openstef-meta==4.2.2",
+  "openstef-models[xgb-cpu]==4.2.2",
 ]
 optional-dependencies.core = [
-    "openstef-core==4.1.2",
+    "openstef-core==4.2.2",
 ]
 urls.Documentation = "https://openstef.github.io/openstef/index.html"
 urls.Homepage = "https://lfenergy.org/projects/openstef/"